page 1  (4 pages)
2to next section

Copyright ? 1996 by L. Hamey and J. Yeh

Appears in Image Segmentation Workshop, pp 65-68,
Australian Pattern Recognition Society, 1996.

Segmentation of Bake Images by a Self-Organising Map

Leonard G. C. Hamey and Jeffrey C.-H. Yeh

Department of Computing

Macquarie University NSW 2109

Phone: (02)9850-9527

Email: [email protected]


A technique for segmentation of images of baked good is presented. The technique employs a SelfOrganising Map to identify the characteristic colour development curve (bake curve) for each product. Segmentation is based upon the colour information contained in the bake curve. The technique is trained with only positive exemplars of the product.


Quality assessment is an important aspect of baked good production. Techniques of manual assessment are prone to errors due to human factors. Previously, we have reported a system for colour baked goods inspection using a combination of the SelfOrganising Map (SOM) [2] and feed-forward neural networks to assess product quality in the manner of human experts but with improved performance [6, 7, 5, 4]. The present paper discusses the image segmentation capabilities of the SOM used in the first stage processing of this quality assessment system.


A key discovery of our work is the existence of the baking curve, a colour development curve obtained by imaging product samples at various levels of bake (figure 1(b)). The curve characterises the colour changes that take place during the baking process. Individual products have distinct baking curves, but different samples of the same product will have colours distributed along the same characteristic curve. This means that product samples in images can be segmented from background distractions on the basis of colour distribution relative to the baking curve.

An 1-dimensional SOM is trained to identify the baking curve. The training data consists only of images of product samples; no examples of nonproduct are required.
A baking curve is usually charcterised by a

sausage shape distribution of pixels in the RGB colour cube. The distribution has a dense centre of biscuit colour pixels and with noise producing a scattering of pixels around this core. A baking line must be therefore extracted along the centre of the baking curve to represent the true path of the colour development during bake. This can be achieved by utilising the properties of a onedimensional SOM.

A SOM is characterised by its topological preserving and equiprobable features. The topological preserving feature ensures that when the trained SOM nodes are projected onto the RGB colour cube, the topology, or baking line, created by connecting neighbouring nodes resembles the shape of the baking curve. The equiprobable feature ensures that the distribution of trained SOM nodes matches the distribution of the pixels in the RGB colour space. Therefore, a SOM with a limited number of nodes will position nodes in the denser centre of the baking curve. This ensures the one-dimensional baking line created lies along the centre of the baking curve and hence correctly represents the baking curve.

Two problems confront the use of the SOM in this application: the selection of the number of nodes and the training duration. We found that a SOM with insufficent nodes will not cover the ends of the baking curve because of the low frequency of pixels representing extremes of bake. On the other hand, a SOM with excessive nodes will start to position the nodes amongst the noise scattered around the baking curve. These two parameters can be determined using a trial-and-error approach with visual analysis.

We collected 298 Milk Coffee biscuit samples of roughly equal distributions for under, correct and over baked categories. Since the bake levels are fairly uniform across the surface of each sample, a 10 by 20 pixels section from each sample was cropped out manually to reduce the size of the training set. This produced a training set of 59600 pixels or training input patterns for the SOM. We trained the SOMs using SOM_PAK, The SelfOrganizing Map Program Package V3.1 [3] which